FAKE NEWS DETECTION USING NATURAL LANGUAGE PROCESSING AND MACHINE LEARNING: A MULTI-DATASET COMPARATIVE STUDY WITH SMOTE-BASED CLASS BALANCING AND HYPERPARAMETER OPTIMIZATION
Keywords:
Fake News Detection, Natural Language Processing, Machine Learning, SMOTE, TF-IDF, Hyperparameter Tuning.Abstract
Unfortunately, fake news has also become an issue in the digital era as it can be damaging to people, organisations and even countries. In this study, a holistic machine learning method to detect fake news using natural language processing (NLP) techniques is presented. The classifiers that are tested are the following: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Passive Aggressive Classifier (PAC), Support Vector Machine (SVM), Multinomial Naive Bayes (MNB). The technique consists of text preprocessing, TF-IDF vectorization, SMOTE (Synthetic Minority Over-sampling Technique) to overcome the class imbalance and Hyperparameter tuning using GridSearchCV and 5-fold cross validation.
We conduct tests with two datasets: a small news dataset (6,335 samples) and a large dataset of mixed fake-real (44,898 samples). With SMOTE and hyperparameter tuning, Naïve Bayes is improved by 5.5% (from 86.0% to 91.5%) and Random Forest by 0.6% (from 90.4% to 91.0%). All of the models achieve a greater accuracy on the larger data set.
